Movie rental by mail (and now by streaming video) company Netflix has been celebrated for its disruptive business model. Woe to Blockbuster, Hollywood Video and the other industry dinosaurs that have regressed to slow follower status, lumbering behind Netflix’s innovation.
That innovation has included the company’s lauded Cinematch function. Cinematch is the Netflix version of Amazon’s purchase circles, a recommendation engine that predicts what you might like to buy next. Cinematch is how Netflix knows that because I like the movie “Almost Famous” I will also like the movie “Lords of Dogtown.” (I did.) Somewhat counter-intuitively, it also knew that I would like the movie “Rabbitproof Fence.” (I did.) Cinematch doesn’t really care what people in my neighborhood are watching, which is fine by me since most of my neighbors are twice my age and tend toward the black-and-white whodunit genre. Of course, I’m guessing on that one, which is my point: Cinematch knows.
But according to Netflix, Cinematch could be even better. The Netflix Prize offers $1 million to the first person or team that can improve Cinematch’s current accuracy by 10 percent or better. The contest was announced on October 26, 2006, and no one has won yet. (So far a team called BellKor in BigChaos is in the lead, delivering a 9.63 percent improvement.)
Why is the problem so vexing? We’ll leave discussions of collaborative filtering, linear regression algorithms and singular value decomposition for another time. One of the most buzzed-about reasons is the so-called Napoleon Dynamite problem. The movie is an outlier: it’s hard to categorize into a single genre. Moreover people either love it or hate it—“Napoleon Dynamite” gets a disproportionate number of 1-star and 5-star ratings. Such a wide preference disparity for a film with such a singular concept can flummox the cleverest predictive algorithms.
The fact is that Cinematch is bound by a singular limitation: a dearth of heterogeneous and comprehensive data. Cinematch doesn’t rely heavily on geography, age, or gender to predict what movies you’ll love, indeed it strips out personal details. Instead it uses information about other movies you’ve liked in the past. The smaller the number of descriptive attributes, the smarter the data mining engine has to be.
Our efforts to understand customer behaviors and preferences are easier. Or at least, they should be if we’re collecting a diverse set of granular-level data about our customers. And we need to standardize and cleanse that data so that it’s consistent. The more information we have about our customer and our products, the more we’ll know about them. That is, the better we’ll understand clear determinants of behaviors and preferences, and craft messages and conversations around them. .
Do the math and it’s clear that upping Cinematch scores is worth vastly more than $1 million to Netflix. What’s the ROI for better customer knowledge at your company? If you don’t know, then you don’t know what you should be spending, and odds are you’re probably not spending enough. Of course I don’t know the precise likelihood that you’re not spending enough, that would call for some analysis. Or we could just go to the movies.
